Pith. sign in

REVIEW 4 major objections 5 minor 69 references

PixelSR: Efficient Screen Content Super-Resolution via Pixel Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read PixelSR claims that classifying each target pixel as unique, repeated, or background lets a screen-content super-resolution network run faster without sacrificing accuracy, while a soft pixel-binning scheme in training adds a content-depend

desk verdict Solid empirical paper with a genuinely new on-the-fly lookup table for arbitrary-scale screen-content SR; the 'without performance loss' claim is plausible but not rigorously proven, and the closest prior comparison is missing. read the letter →

arxiv 2608.00646 v1 pith:2D4LRUBP submitted 2026-08-01 cs.CV

Zhiheng Li , Lei Chen , Jie Zhou , Jiwen Lu This is my paper
classification cs.CV
keywords screencontentsuper-resolutionpixelclassificationsoftbinningattentionon-the-flylookuptableimplicitneuralrepresentationefficientinferencescale-arbitrarySR
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that screen content images—texts, charts, tables—have two properties natural images lack: unnaturally sharp edges and heavy pixel-level repetition. PixelSR exploits these in both phases. In training, a lightweight content extractor softly bins pixels into content groups and pools features within each group to build a content attention map, giving every pixel a content-dependent, non-local receptive field. In inference, it divides target HR pixels into unique, repeated, and background classes: unique pixels are processed by the network and cached, repeated pixels are fetched from an on-the-fly lookup table, and flat background pixels are filled by nearest-neighbor. The claim is that this improves PSNR over prior screen-content SR methods (e.g., 38.21 vs 38.02 dB at x2 and 33.52 vs 32.80 at x3 on SCI1K with EDSR backbone) while shortening total inference time, without the accuracy-speed tradeoff that earlier efficient-SR schemes accepted.

What carries the argument

The two load-bearing mechanisms are (1) soft pixel binning with content attention in training and (2) the on-the-fly lookup table plus nearest-neighbor in inference. Soft binning assigns each pixel soft weights across N=10 uniformly spaced bins via a temperature-scaled exponential of distance to bin centers, preserving gradient flow; bins are merged into two content groups, and group-mean features are pooled spatially and broadcast back, yielding a content-dependent, non-local receptive field for every pixel. In inference, HR pixels are keyed by the RGB values of the nearest 3x3 LR patch plus the relative HR coordinate; the on-the-fly lookup table caches outputs for unique pixels and serves

What would settle it

Measure, on the SCI1K test set, how often the on-the-fly lookup table returns a prediction that differs from running the full network on the same 'repeated' pixel inputs, and compute the resulting PSNR difference; if the mismatch rate is non-negligible or the PSNR gap exceeds the reported fluctuation bounds, the lookup-table equivalence claim fails. A simpler version: count the fraction of repeated pixels whose cached key collides with a unique pixel but whose encoder feature (before the decoder) is not identical.

Watch

Extended reading notes

Core claim

PixelSR's central discovery is that pixel-level classification, rather than window or sub-image classification, is the right granularity for efficient screen-content SR. During training, soft binning discretizes pixels into two content-dependent groups—sharp screen-content edges vs. the rest—and spatial average pooling of each group yields a content attention vector that is broadcast back to every pixel, effectively enlarging the receptive field beyond the local 3x3 patch used by the implicit neural representation decoder. During inference, the method exploits self-repetitiveness and flatness: targeted HR pixels whose nearest 3x3 LR patch has equal RGB values are background pixels handled by

Load-bearing premise

The speedup collapses if two pixels with the same 3x3 local patch and the same relative coordinate do not reliably get the same network output—which the paper assumes without measuring, while the encoder's global receptive field could in principle make distant context change the prediction.

Editorial extensions

If this is right

  • PixelSR achieves state-of-the-art PSNR on SCI1K, SCID, and SIQAD at in-training scales (e.g., 33.52 dB vs 32.80 dB at x3 on SCI1K with EDSR backbone) and maintains the lead at out-of-training scales up to x10.
  • Total inference time on the SCI1K test set is shorter than LIIF, ITSRN, LTE, and BTC at every tested scale factor, with no PSNR drop attributable to the lookup table.
  • The on-the-fly lookup table is a plug-in accelerator: the paper shows it reduces inference time for LIIF, ITSRN, LTE, BTC, and PixelSR alike, with PSNR fluctuations within small values.
  • The content attention from soft binning gives a content-dependent and unlimited receptive field, which is the mechanism behind the accuracy gains; ablation shows 10 bins and a single conv layer are the best settings.
  • The method generalizes across backbones (EDSR-baseline and SwinIR-light) and to other screen-content benchmarks, indicating the classification scheme, not extra capacity, drives the gains (all models ~1.6-1.9M params).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the lookup-table key is truly sufficient, the same trick should transfer to other content with repetitive structure—e.g., UI screenshots, CAD drawings, or document scans—suggesting a testable extension: measure the speedup and PSNR on such corpora to see where the repetition statistics break down.
  • A direct falsifier is suggested by the paper's own design: because the encoder backbone has a global receptive field through downsampling layers, two pixels with identical 3x3 local patches and relative offsets can still receive different encoder features; one could quantify mismatch frequency and show whether errors accumulate as scale increases.
  • The repetition statistics (20.90% background, 42.25% repeated patches in SCI1K vs 2.08% and 3.51% in DIV2K) imply the speedup is a property of the dataset, not the model; a natural next step is content-adaptive routing that decides per patch whether to run the network or the lookup table, making the method robust to mixed content.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes PixelSR, a screen-content super-resolution method with two components. In training, a lightweight content extractor plus soft pixel binning splits pixels into content-dependent groups; spatial average pooling within each group yields a content attention map that is concatenated to encoder features, giving every pixel a global, content-dependent receptive field. In inference, targeted HR pixels are classified as unique, repeated, or background. Unique pixels are processed by the network and their outputs cached in an on-the-fly lookup table keyed by the 3×3 LR patch RGB values and the relative HR coordinate; repeated pixels retrieve cached outputs; background pixels are generated by an unspecified nearest-neighbor scheme. Experiments on SCI1K, SCID, and SIQAD with EDSR and SwinIR-light backbones claim state-of-the-art PSNR at scales ×2–×10 and shorter total inference time on SCI1K without performance loss.

Significance. If the claims hold, PixelSR would be a meaningful contribution: it directly exploits two distinctive properties of screen content (uniform backgrounds and self-repetition) for both accuracy and speed, and it generalizes the caching idea to scale-arbitrary SR. The empirical breadth is a strength: three datasets, two backbone families, scales up to ×10, and comparisons to four strong baselines. The statistical evidence that background and repeated 3×3 patches are far more frequent in SCI1K than in DIV2K is a useful motivation. However, the central 'no performance loss' claim for the lookup table is not rigorously established, and one reported result contradicts the text. The paper currently does not ship code or machine-checked proofs, so the reproduction of the speedup and PSNR comparisons depends on the completeness of the experimental description.

major comments (4)
  1. [Sec. III-B, Table III] The central claim of 'faster inference without performance loss' rests on the on-the-fly lookup table's assumption that two pixels with the same (3×3 LR patch, relative coordinate) key receive identical network outputs. This is not guaranteed by the architecture: the EDSR/SwinIR encoders have receptive fields larger than 3×3, and F_CA in Eq. (6) is computed from whole-image group pooling (Eqs. 4–5), so the decoder input F_cat depends on distant context. Two pixels sharing the same local key can therefore produce different outputs. Table III only reports aggregate PSNR differences of 0.01–0.03 dB, with no per-pixel mismatch statistics, error bars, or significance test. Please add a direct measurement: run the full network for all pixels, compare predictions for repeated pixels against unique pixels with identical keys, and report the mismatch rate and its PSNR contribution. Alternatively,
  2. [Sec. IV-B, Table I] The text states that 'PixelSR consistently achieves the highest PSNR across all test sets at all scale factors' and 'ranks first across all scales' on SIQAD. Table I shows that on SIQAD at ×2, LTE achieves 33.29 dB while PixelSR achieves 33.25 dB. This contradicts the stated claim and affects the 'state-of-the-art' assertion. Please correct the claim or the table, and re-check every 'best' marker and the corresponding narrative.
  3. [Sec. IV-A / Sec. III-C] Reproducibility is hindered by missing settings and algorithmic details. The soft-binning temperature τ in Eq. (3) is never given a value, despite being a free parameter that controls the hardness of binning. The background nearest-neighbor algorithm is described only as 'use the nearest neighbor algorithm to generate HR pixels' without specifying the neighbor source, distance metric, or how it is applied at HR. These details are needed to reproduce the reported speed and quality numbers.
  4. [Table III vs. Table I] There is a numerical inconsistency: ITSRN without the lookup table at ×5 is reported as 26.52 dB in Table III but 26.72 dB in Table I. The inference-time methodology is also under-specified: what GPU, batch size, number of runs, and whether the timing includes the backbone encoder? The statement that 'inference time fluctuates within 1 second' is insufficient to assess the reported speedups.
minor comments (5)
  1. [Sec. IV-B] The text reports 'PixelSR achieves 28.54 dB on SCID at ×3 scale' with the SwinIR backbone, but Table II lists 28.59 dB; please harmonize.
  2. [Sec. III-B, Fig. 3] The statistical comparison of background and repeated patches should specify whether 3×3 patches are extracted with stride 1 or non-overlapping. The percentages differ substantially depending on this choice.
  3. [Sec. III-B] The statement that the on-the-fly lookup table is 'the first lookup table that can be used for scale arbitrary SR' is not substantiated by a comparison to prior arbitrary-scale LUT work; please soften or provide supporting references.
  4. [References] Some references are duplicated (e.g., [10]/[17], [5]/[18]) and some venue names are inconsistent (e.g., 'NeurIPS' vs. full conference name). Please clean the bibliography.
  5. [Abstract / Contributions] The phrase 'without the loss of model performance' should be qualified as 'with negligible PSNR difference on aggregate' unless per-pixel equivalence is proven; the current wording overstates the evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core PSNR and speed claims rest on a new architecture evaluated against external baselines; the lookup-table equivalence issue is a correctness/validity concern, not a circular reduction.

full rationale

The paper's derivation chain is self-contained. The content-attention module is a new feedforward computation defined by Eqs. 1-6, trained with L1 loss, and compared against external published methods (LIIF, ITSRN, LTE, BTC) on public screen-content benchmarks. No load-bearing step reduces to a fitted parameter renamed as a prediction. The number of bins, temperature, and group split are hyperparameters selected via ablations on the test set (Tables IV/V), which is a methodological weakness (test-set selection) but not a definitional circularity. The inference lookup table is an approximate caching scheme: the paper reports PSNR with and without the table (Table III) and finds small differences (0.00-0.03 dB). The claim that the 3x3 LR patch plus relative offset is a sufficient key is not guaranteed by the network's global receptive field, and no per-pixel mismatch statistics are provided; this is an unsupported equivalence and a validity risk, not a circular step, because the table output is not used to define the network output. The paper itself notes in V. Conclusion that the on-the-fly lookup table 'would not function well in natural images,' a scope limitation rather than a circularity. Self-citations in the references (e.g., [8], [9], [62]) are contextual and not load-bearing. Therefore, no significant circularity is identified.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method introduces no new physical entities. Its free parameters are architectural hyperparameters (bin count, temperature, group split, extractor depth), of which bin count and layer count are ablated on the test set; temperature and group split are not. The load-bearing domain assumptions are the patch-equivalence key for the lookup table and the flatness of background pixels, both untested against the encoder's global receptive field and real degradation.

free parameters (4)
  • bin temperature tau = not reported
    Eq. 3: temperature controlling softness of bin assignment; chosen by hand as 'low', no reported value or sensitivity analysis.
  • number of bins N = 10
    Section III.C sets 10 bins; ablation Table IV shows PSNR varies with bin count (2/5/8/10), so it is a tuned hyperparameter.
  • group split P = P=1 (first bin vs rest)
    Section III.C: first bin with lowest center is group 1, rest 9 bins merged as group 2; chosen by design without an ablation.
  • content extractor layer count/kernel = 1 layer, 3x3 kernel
    Ablated in Table V; 1 layer chosen as best, but the choice is validated on the test set.
assumptions (3)
  • domain assumption Exact equality of 3x3 RGB patch + relative distance implies (nearly) identical network output for the target pixel
    Used by the on-the-fly lookup table (Section III.B, Fig. 4). True only if encoder features are also identical; the encoder has receptive field larger than 3x3, so the assumption is approximate and unverified.
  • domain assumption Background pixels (flat 3x3 LR patch) are best upscaled by nearest-neighbor, i.e., their HR neighborhood is also flat
    Section III.B: nearest-neighbor for background pixels. Valid for ideal screen content but can break at anti-aliased or compressed edges; no error analysis given.
  • domain assumption Screen-content datasets are representative of the deployment distribution and bicubic downsampling is the right degradation model
    Training/eval protocol (Section IV.A): bicubic downsampling with scales 1-4; all comparisons are on that synthetic degradation. Real screen-content transmission artifacts (compression, chroma subsampling) are not tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PixelSR: Efficient Screen Content Super-Resolution via Pixel Classification." pith.science (2026). https://pith.science/paper/2D4LRUBP

@misc{pith2026260800646,
  author       = {Pith},
  title        = {Pith review of: PixelSR: Efficient Screen Content Super-Resolution via Pixel Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2D4LRUBP}},
  note         = {Machine review of arXiv:2608.00646}
}
read the original abstract

Screen content images are generally composed of texts and graphics. Compared to natural images, these man-made images contain a large quantity of sharp but repetitive structures. However, existing works in screen content super-resolution underutilize the special characteristics of screen content, leaving a large room to improve model performance and speed up. In this paper, we propose PixelSR, a simple yet effective method to improve super-resolution performance but with faster inference speed. To improve model performance, we classify pixels via pixel binning to compute content attention in the training phase. Specifically, after binning pixels into content-dependent groups, content attention is aggregated from pixel features within each group to introduce a content-dependent and non-local receptive field for every pixel. In the testing phase, we utilize the properties of self-repetitiveness and redundancy in screen content to speed up inference without the loss of model performance. We divide targeted high-resolution pixels into three types, which are unique pixels, repeated pixels, and background pixels for each test image. We conduct conventional network processing on unique pixels and cache their predictions in the on-the-fly lookup table. For repeated pixels which have appeared in unique pixels, we directly retrieve prediction results from the lookup table without network processing. For background pixels, we use the nearest neighbor algorithm to generate high-resolution pixels. The on-the-fly lookup table is cleaned and repeats the procedure above for the next test image. Experiments show our PixelSR achieves state-of-the-art performance with shorter inference time in screen content super-resolution.

Figures

Figures reproduced from arXiv: 2608.00646 by the authors.

Figure 1
Figure 1. Comparison between PixelSR and conventional efficient SR paradigm: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The training pipeline of PixelSR. PixelSR uses a content extractor to generate a score map [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. (a) The example of screen content image; (b) the corresponding image where the background patches are masked out (in black) and the repeated [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The inference pipeline of PixelSR. For each test image, PixelSR divides pixels into unique pixels, repeated pixels, and background pixels. PixelSR [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison between different methods at the scale factors of 3 and 5. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Further qualitative comparison between methods for the scale factor of 4 with the scene text recognition (STR) results by a pretrained STR network [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visualization of the pixel binning result for each input. The edges of screen content such as texts and charts are classified into the second group [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Comparison of inference time (excluding the running time of the [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Visual Comparison of the pixels classified into the second group for the number of bins of (b) 2, (c) 5, (d) 8, and (e) 10 (PixelSR) during training. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Visual Comparison of the pixels classified into the second group for the number of layers of (b) 3, (c) 2, and (d) 1 (PixelSR) during training. It [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 3 linked inside Pith

  1. [30]

    Accelerating image super- resolution networks with pixel-level classification,

    J. Jeong, J. Kim, Y . Jo, and S. J. Kim, “Accelerating image super- resolution networks with pixel-level classification,”arXiv preprint arXiv:2407.21448, 2024. 1, 3

    arXiv 2024

  2. [1]

    Deep learning for single image super-resolution: A brief review,

    W. Yang, X. Zhang, Y . Tian, W. Wang, J.-H. Xue, and Q. Liao, “Deep learning for single image super-resolution: A brief review,”TMM, vol. 21, no. 12, pp. 3106–3121, 2019. 1

    2019

  3. [2]

    Single-image super-resolution: A benchmark,

    C.-Y . Yang, C. Ma, and M.-H. Yang, “Single-image super-resolution: A benchmark,” inECCV, 2014, pp. 372–386. 1

    2014

  4. [3]

    Transformer for single image super-resolution,

    Z. Lu, J. Li, H. Liu, C. Huang, L. Zhang, and T. Zeng, “Transformer for single image super-resolution,” inCVPR, 2022, pp. 457–466. 1

    2022

  5. [4]

    Single image super-resolution from transformed self-exemplars,

    J.-B. Huang, A. Singh, and N. Ahuja, “Single image super-resolution from transformed self-exemplars,” inCVPR, 2015, pp. 5197–5206. 1

    2015

  6. [6]

    Single-image super-resolution using sparse regression and natural image prior,

    K. I. Kim and Y . Kwon, “Single-image super-resolution using sparse regression and natural image prior,”TPAMI, vol. 32, no. 6, pp. 1127– 1133, 2010. 1

    2010

  7. [7]

    Second-order attention network for single image super-resolution,

    T. Dai, J. Cai, Y . Zhang, S.-T. Xia, and L. Zhang, “Second-order attention network for single image super-resolution,” inCVPR, 2019, pp. 11 065–11 074. 1

    2019

  8. [8]

    Structure-preserving image super- resolution,

    C. Ma, Y . Rao, J. Lu, and J. Zhou, “Structure-preserving image super- resolution,”TPAMI, vol. 44, no. 11, pp. 7898–7911, 2021. 1

    2021

Show all 69 references
  1. [9]

    Structure- preserving super resolution with gradient guidance,

    C. Ma, Y . Rao, Y . Cheng, C. Chen, J. Lu, and J. Zhou, “Structure- preserving super resolution with gradient guidance,” inCVPR, 2020, pp. 7769–7778. 1

    2020

  2. [10]

    Limits on super-resolution and how to break them,

    S. Baker and T. Kanade, “Limits on super-resolution and how to break them,”TPAMI, vol. 24, no. 9, pp. 1167–1183, 2002. 1

    2002

  3. [11]

    Deep networks for image super-resolution with sparse prior,

    Z. Wang, D. Liu, J. Yang, W. Han, and T. Huang, “Deep networks for image super-resolution with sparse prior,” inICCV, 2015, pp. 370–378. 1

    2015

  4. [12]

    Multi-scale residual network for image super-resolution,

    J. Li, F. Fang, K. Mei, and G. Zhang, “Multi-scale residual network for image super-resolution,” inECCV, 2018, pp. 517–532. 1

    2018

  5. [13]

    Deep burst super- resolution,

    G. Bhat, M. Danelljan, L. Van Gool, and R. Timofte, “Deep burst super- resolution,” inCVPR, 2021, pp. 9209–9218. 1

    2021

  6. [14]

    Deep learning for image super- resolution: A survey,

    Z. Wang, J. Chen, and S. C. Hoi, “Deep learning for image super- resolution: A survey,”TPAMI, vol. 43, pp. 3365–3387, 2020. 1

    2020

  7. [15]

    A survey on super-resolution imaging,

    J. Tian and K.-K. Ma, “A survey on super-resolution imaging,”SIGNAL IMAGE VIDEO P, vol. 5, pp. 329–342, 2011. 1 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

    2011

  8. [16]

    Anycost network quantization for image super-resolution,

    J. Zhang, Z. Wang, H. Wang, J. Zhou, and J. Lu, “Anycost network quantization for image super-resolution,”TIP, vol. 33, pp. 2279–2292,

  9. [17]

    Limits on super-resolution and how to break them,

    S. Baker and T. Kanade, “Limits on super-resolution and how to break them,”IEEE transactions on pattern analysis and machine intelligence, vol. 24, pp. 1167–1183, 2002. 1

    2002

  10. [18]

    Super-resolution from a single image,

    D. Glasner, S. Bagon, and M. Irani, “Super-resolution from a single image,” inICCV, 2009, pp. 349–356. 1

    2009

  11. [19]

    Space-time super-resolution,

    E. Shechtman, Y . Caspi, and M. Irani, “Space-time super-resolution,” TPAMI, vol. 27, pp. 531–545, 2005. 1

    2005

  12. [20]

    Learning a deep convolutional network for image super-resolution,

    C. Dong, C. C. Loy, K. He, and X. Tang, “Learning a deep convolutional network for image super-resolution,” inECCV, 2014, pp. 184–199. 1

    2014

  13. [21]

    Image super-resolution using deep convolutional networks,

    ——, “Image super-resolution using deep convolutional networks,” TPAMI, vol. 38, pp. 295–307, 2015. 1

    2015

  14. [22]

    Statistical performance analysis of super- resolution,

    D. Robinson and P. Milanfar, “Statistical performance analysis of super- resolution,”IEEE Transactions on Image Processing, vol. 15, pp. 1413– 1428, 2006. 1

    2006

  15. [23]

    Bayesian image super-resolution,

    M. Tipping and C. Bishop, “Bayesian image super-resolution,”NeurIPS, vol. 15, 2002. 1

    2002

  16. [24]

    Learning texture transformer network for image super-resolution,

    F. Yang, H. Yang, J. Fu, H. Lu, and B. Guo, “Learning texture transformer network for image super-resolution,” inCVPR, 2020, pp. 5791–5800. 1

    2020

  17. [25]

    Masa-sr: Matching acceleration and spatial adaptation for reference-based image super-resolution,

    L. Lu, W. Li, X. Tao, J. Lu, and J. Jia, “Masa-sr: Matching acceleration and spatial adaptation for reference-based image super-resolution,” in CVPR, 2021, pp. 6368–6377. 1

    2021

  18. [26]

    B-spline texture coefficients estimator for screen content image super-resolution,

    B. Pak, J. Lee, and K. H. Jin, “B-spline texture coefficients estimator for screen content image super-resolution,” inCVPR, 2023, pp. 10 062– 10 071. 1, 2, 3, 6, 7, 8, 9, 10, 12

    2023

  19. [27]

    Perceptual quality assessment of screen content images,

    H. Yang, Y . Fang, and W. Lin, “Perceptual quality assessment of screen content images,”TIP, no. 11, pp. 4408–4421, 2015. 1, 3, 6

    2015

  20. [28]

    Classsr: A general framework to accelerate super-resolution networks by data characteristic,

    X. Kong, H. Zhao, Y . Qiao, and C. Dong, “Classsr: A general framework to accelerate super-resolution networks by data characteristic,” inCVPR, 2021, pp. 12 016–12 025. 1, 3

    2021

  21. [29]

    Camixersr: Only details need more

    Y . Wang, S. Zhao, Y . Liu, J. Li, and L. Zhang, “Camixersr: Only details need more" attention",”arXiv preprint arXiv:2402.19289, 2024. 1, 3

    2024 arXiv

  22. [31]

    Ciaosr: Continuous implicit attention-in- attention network for arbitrary-scale image super-resolution,

    J. Cao, Q. Wang, Y . Xian, Y . Li, B. Ni, Z. Pi, K. Zhang, Y . Zhang, R. Timofte, and L. Van Gool, “Ciaosr: Continuous implicit attention-in- attention network for arbitrary-scale image super-resolution,” inCVPR, 2023, pp. 1796–1807. 1, 2, 3

    2023

  23. [32]

    Image super-resolution with cross-scale non-local attention and exhaustive self- exemplars mining,

    Y . Mei, Y . Fan, Y . Zhou, L. Huang, T. S. Huang, and H. Shi, “Image super-resolution with cross-scale non-local attention and exhaustive self- exemplars mining,” inCVPR, 2020, pp. 5690–5699. 1, 3

    2020

  24. [33]

    Image processing gnn: Breaking rigidity in super-resolution,

    Y . Tian, H. Chen, C. Xu, and Y . Wang, “Image processing gnn: Breaking rigidity in super-resolution,” inCVPR, 2024, pp. 24 108–24 117. 1, 3

    2024

  25. [34]

    Learning continuous image representa- tion with local implicit image function,

    Y . Chen, S. Liu, and X. Wang, “Learning continuous image representa- tion with local implicit image function,” inCVPR, 2021, pp. 8628–8638. 2, 3, 6, 7, 8, 9, 10, 12

    2021

  26. [35]

    Implicit transformer network for screen content image continuous super-resolution,

    J. Yang, S. Shen, H. Yue, and K. Li, “Implicit transformer network for screen content image continuous super-resolution,”NeurIPS, pp. 13 304– 13 315, 2021. 2, 3, 5, 6, 7, 8, 9, 10, 12

    2021

  27. [36]

    Local texture estimator for implicit representation function,

    J. Lee and K. H. Jin, “Local texture estimator for implicit representation function,” inCVPR, 2022, pp. 1929–1938. 2, 3, 7, 8, 9, 10, 12

    2022

  28. [38]

    Learning a single network for scale-arbitrary super-resolution,

    L. Wang, Y . Wang, Z. Lin, J. Yang, W. An, and Y . Guo, “Learning a single network for scale-arbitrary super-resolution,” inICCV, 2021, pp. 4801–4810. 2

    2021

  29. [39]

    Deep arbitrary- scale image super-resolution via scale-equivariance pursuit,

    X. Wang, X. Chen, B. Ni, H. Wang, Z. Tong, and Y . Liu, “Deep arbitrary- scale image super-resolution via scale-equivariance pursuit,” inCVPR, 2023, pp. 1786–1795. 2

    2023

  30. [40]

    Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution,

    X. Xu, Z. Wang, and H. Shi, “Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution,” arXiv preprint arXiv:2103.12716, 2021. 2

    2021 arXiv

  31. [41]

    Local implicit normalizing flow for arbitrary-scale image super- resolution,

    J.-E. Yao, L.-Y . Tsao, Y .-C. Lo, R. Tseng, C.-C. Chang, and C.-Y . Lee, “Local implicit normalizing flow for arbitrary-scale image super- resolution,” inCVPR, 2023, pp. 1776–1785. 2

    2023

  32. [42]

    Cascaded local implicit transformer for arbitrary-scale super- resolution,

    H.-W. Chen, Y .-S. Xu, M.-F. Hong, Y .-M. Tsai, H.-K. Kuo, and C.- Y . Lee, “Cascaded local implicit transformer for arbitrary-scale super- resolution,” inCVPR, 2023, pp. 18 257–18 267. 2

    2023

  33. [43]

    Activating more information in arbitrary-scale image super-resolution,

    Y . Zhao, Q. Teng, H. Chen, S. Zhang, X. He, Y . Li, and R. E. Sheriff, “Activating more information in arbitrary-scale image super-resolution,” TMM, vol. 26, pp. 7946–7961, 2024. 2

    2024

  34. [44]

    Recovering realistic details for magnification-arbitrary image super-resolution,

    C. Ma, P. Yu, J. Lu, and J. Zhou, “Recovering realistic details for magnification-arbitrary image super-resolution,”TIP, vol. 31, pp. 3669– 3683, 2022. 2

    2022

  35. [45]

    Hiif: Hierarchical encoding based implicit image function for continuous super-resolution,

    Y . Jiang, H. M. Kwan, T. Peng, G. Gao, F. Zhang, X. Zhu, J. Sole, and D. Bull, “Hiif: Hierarchical encoding based implicit image function for continuous super-resolution,” inCVPR, 2025, pp. 2289–2299. 2

    2025

  36. [46]

    Meta-sr: A magnification-arbitrary network for super-resolution,

    X. Hu, H. Mu, X. Zhang, Z. Wang, T. Tan, and J. Sun, “Meta-sr: A magnification-arbitrary network for super-resolution,” inCVPR, 2019, pp. 1575–1584. 2

    2019

  37. [47]

    Learning based screen image compres- sion,

    H. Yang, W. Lin, and C. Deng, “Learning based screen image compres- sion,” inMMSPW, 2012, pp. 77–82. 3

    2012

  38. [48]

    Super resolution for compressed screen content video,

    M. Wang, J. Xu, L. Zhang, J. Li, and S. Wang, “Super resolution for compressed screen content video,” inDCC, 2021, pp. 173–182. 3

    2021

  39. [49]

    Itsrn++: Stronger and better implicit transformer network for continuous screen content image super- resolution,

    S. Shen, H. Yue, J. Yang, and K. Li, “Itsrn++: Stronger and better implicit transformer network for continuous screen content image super- resolution,”arXiv, 2022. 3

    2022

  40. [50]

    Real-time single image and video super- resolution using an efficient sub-pixel convolutional neural network,

    W. Shi, J. Caballero, F. Huszár, J. Totz, A. P. Aitken, R. Bishop, D. Rueckert, and Z. Wang, “Real-time single image and video super- resolution using an efficient sub-pixel convolutional neural network,” in CVPR, 2016, pp. 1874–1883. 3

    2016

  41. [51]

    Fast, accurate, and lightweight super-resolution with cascading residual network,

    N. Ahn, B. Kang, and K.-A. Sohn, “Fast, accurate, and lightweight super-resolution with cascading residual network,” inECCV, 2018, pp. 252–268. 3

    2018

  42. [52]

    Lightweight image super- resolution with information multi-distillation network,

    Z. Hui, X. Gao, Y . Yang, and X. Wang, “Lightweight image super- resolution with information multi-distillation network,” inACM MM, 2019, pp. 2024–2032. 3

    2019

  43. [53]

    Enhanced deep residual networks for single image super-resolution,

    B. Lim, S. Son, H. Kim, S. Nah, and K. Mu Lee, “Enhanced deep residual networks for single image super-resolution,” inCVPRW, 2017, pp. 136–144. 3, 6, 7

    2017

  44. [54]

    Lattice network for lightweight image restoration,

    X. Luo, Y . Qu, Y . Xie, Y . Zhang, C. Li, and Y . Fu, “Lattice network for lightweight image restoration,”TPAMI, no. 4, pp. 4826–4842, 2022. 3

    2022

  45. [55]

    Dlgsanet: lightweight dynamic local and global self-attention networks for image super-resolution,

    X. Li, J. Dong, J. Tang, and J. Pan, “Dlgsanet: lightweight dynamic local and global self-attention networks for image super-resolution,” in ICCV, 2023, pp. 12 792–12 801. 3

    2023

  46. [56]

    Accelerating the super-resolution convolutional neural network,

    C. Dong, C. C. Loy, and X. Tang, “Accelerating the super-resolution convolutional neural network,” inECCV, 2016, pp. 391–407. 3

    2016

  47. [57]

    Ntire 2023 challenge on efficient super- resolution: Methods and results,

    Y . Li, Y . Zhang, R. Timofte, L. Van Gool, L. Yu, Y . Li, X. Li, T. Jiang, Q. Wu, M. Hanet al., “Ntire 2023 challenge on efficient super- resolution: Methods and results,” inCVPR, 2023, pp. 1921–1959. 3

    2023

  48. [58]

    Global aligned structured sparsity learning for efficient image super-resolution,

    H. Wang, Y . Zhang, C. Qin, L. Van Gool, and Y . Fu, “Global aligned structured sparsity learning for efficient image super-resolution,”TPAMI,

  49. [59]

    Lmlt: Low-to-high multi-level vision transformer for lightweight image super-resolution,

    J. Kim, J. Nang, and J. Choe, “Lmlt: Low-to-high multi-level vision transformer for lightweight image super-resolution,” inICCV, 2025, pp. 5509–5519. 3

    2025

  50. [60]

    Practical single-image super-resolution using look- up table,

    Y . Jo and S. J. Kim, “Practical single-image super-resolution using look- up table,” inCVPR, 2021, pp. 691–700. 3

    2021

  51. [61]

    Mulut: Cooperating multiple look-up tables for efficient image super-resolution,

    J. Li, C. Chen, Z. Cheng, and Z. Xiong, “Mulut: Cooperating multiple look-up tables for efficient image super-resolution,” inECCV, 2022, pp. 238–256. 3

    2022

  52. [62]

    Learning series-parallel lookup tables for efficient image super-resolution,

    C. Ma, J. Zhang, J. Zhou, and J. Lu, “Learning series-parallel lookup tables for efficient image super-resolution,” inECCV, 2022, pp. 305–

    2022

  53. [63]

    Reconstructed convolution module based look-up tables for efficient image super- resolution,

    G. Liu, Y . Ding, M. Li, M. Sun, X. Wen, and B. Wang, “Reconstructed convolution module based look-up tables for efficient image super- resolution,” inICCV, 2023, pp. 12 217–12 226. 3

    2023

  54. [64]

    Arm: Any-time super-resolution method,

    B. Chen, M. Lin, K. Sheng, M. Zhang, P. Chen, K. Li, L. Cao, and R. Ji, “Arm: Any-time super-resolution method,” inECCV, 2022, pp. 254–270. 3

    2022

  55. [65]

    Cabm: Content- aware bit mapping for single image super-resolution network with large input,

    S. Tian, M. Lu, J. Liu, Y . Guo, Y . Chen, and S. Zhang, “Cabm: Content- aware bit mapping for single image super-resolution network with large input,” inCVPR, 2023, pp. 1756–1765. 3

    2023

  56. [66]

    Autolut: Lut-based image super-resolution with automatic sampling and adaptive residual learning,

    Y . Xu, S. Yang, X. Liu, J. Liu, J. Tang, and G. Wu, “Autolut: Lut-based image super-resolution with automatic sampling and adaptive residual learning,” inCVPR, 2025, pp. 23 131–23 140. 3

    2025

  57. [67]

    Ntire 2017 challenge on single image super-resolution: Dataset and study,

    E. Agustsson and R. Timofte, “Ntire 2017 challenge on single image super-resolution: Dataset and study,” inCVPRW, 2017, pp. 126–135. 5

    2017

  58. [68]

    Esim: Edge similarity for screen content image quality assessment,

    Z. Ni, L. Ma, H. Zeng, J. Chen, C. Cai, and K.-K. Ma, “Esim: Edge similarity for screen content image quality assessment,”TIP, no. 10, pp. 4818–4831, 2017. 6

    2017

  59. [69]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Swinir: Image restoration using swin transformer,” inICCV, 2021, pp. 1833–1844. 6, 7

    2021

  60. [70]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv, 2014. 6

    2014

  61. [71]

    Scene text recognition with permuted autoregressive sequence models,

    D. Bautista and R. Atienza, “Scene text recognition with permuted autoregressive sequence models,” inECCV, 2022, pp. 178–196. 8

    2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.